nx上darknet的使用-目标检测-在python中的使用

1  内置的代码

在darknet中已经内置了两个py文件

darknet_video.py与darknet_images.py用法类似,都是改一改给的参数就行了,我们说一下几个关键的参数

  • input 要预测哪张图像
  • weights 要使用哪个权重
  • config_file 要使用哪个cfg文件
  • data_file 要使用哪个data文件
  • thresh 置信度给多少

改完上面这些参数就可以直接运行了

用鼠标点一下图片然后按q可以关闭图片

调用视频有时可以正常用,有时不可以,会报下面的错,可能是opencv的版本问题

这两个代码都很好,想的很周道,但是看起来太麻烦,我通常用下面两个简化版的

2  opencv使用模型

优点:只要有python,有opencv-python,有numpy就能用(不需要编译darknet,只要有names、cfg、weights就行)

缺点:只能搞CPU版,用摄像头搞效果较差

2.1  识别图像

你需要一张预测图像,训练时的classes.names yolov4-tiny-cfg yolov4-tiny_final.weights

yolo_opencv.py的内容如下

import cv2
import numpy as npLABELS = open("classes.names").read().strip().split("\n")
net = cv2.dnn.readNetFromDarknet('yolov4-tiny.cfg', 'yolov4-tiny_final.weights')
layer = net.getUnconnectedOutLayersNames()frame = cv2.imread('000003.jpg')
(H, W) = frame.shape[:2]
blob = cv2.dnn.blobFromImage(frame, 1 / 255.0, (416, 416),swapRB=True, crop=False)
net.setInput(blob)
layerOutputs = net.forward(layer)
boxes = []
confidences = []
classIDs = []for output in layerOutputs:for detection in output:scores = detection[5:]classID = np.argmax(scores)confidence = scores[classID]box = detection[0:4] * np.array([W, H, W, H])(centerX, centerY, width, height) = box.astype("int")x = int(centerX - (width / 2))y = int(centerY - (height / 2))boxes.append([x, y, int(width), int(height)])confidences.append(float(confidence))classIDs.append(classID)idxs = cv2.dnn.NMSBoxes(boxes, confidences, 0.5, 0.3)
if len(idxs) > 0:for i in idxs.flatten():(x, y) = (boxes[i][0], boxes[i][1])(w, h) = (boxes[i][2], boxes[i][3])cv2.rectangle(frame, (x, y), (x + w, y + h), (0,255,0), 1, lineType=cv2.LINE_AA)text = "{}: {:.4f}".format(LABELS[classIDs[i]], confidences[i])cv2.putText(frame, text, (x, y - 5), cv2.FONT_HERSHEY_SIMPLEX,0.5, (255,0,0), 1, lineType=cv2.LINE_AA)cv2.imshow('frame',frame)
cv2.waitKey(0)
cv2.destroyAllWindows()

运行后可以显示预测结果

2.2  flask起服务

我们可以搞一个接口来处理识别的功能

服务端

import numpy as np
import cv2
from flask import Flask,request
import base64LABELS = open("classes.names").read().strip().split("\n")
net = cv2.dnn.readNetFromDarknet('yolov4-tiny.cfg', 'yolov4-tiny_final.weights')
layer = net.getUnconnectedOutLayersNames()app = Flask(__name__)
@app.route('/predict',methods=["POST"])
def predict():if request.method == 'POST':image_base64 = request.json['img_b64']img = base64.b64decode(image_base64)img = np.fromstring(img,np.uint8)frame = cv2.imdecode(img,cv2.IMREAD_COLOR)(H, W) = frame.shape[:2]blob = cv2.dnn.blobFromImage(frame, 1 / 255.0, (416, 416),swapRB=True, crop=False)net.setInput(blob)layerOutputs = net.forward(layer)boxes = []confidences = []classIDs = []for output in layerOutputs:for detection in output:scores = detection[5:]classID = np.argmax(scores)confidence = scores[classID]box = detection[0:4] * np.array([W, H, W, H])(centerX, centerY, width, height) = box.astype("int")x = int(centerX - (width / 2))y = int(centerY - (height / 2))boxes.append([x, y, int(width), int(height)])confidences.append(float(confidence))classIDs.append(classID)idxs = cv2.dnn.NMSBoxes(boxes, confidences, 0.5, 0.3)result = []if len(idxs) > 0:for i in idxs.flatten():(x, y) = (boxes[i][0], boxes[i][1])(w, h) = (boxes[i][2], boxes[i][3])result_obj = {}result_obj['x'] = xresult_obj['y'] = yresult_obj['w'] = wresult_obj['h'] = hresult_obj['name'] = LABELS[classIDs[i]]result_obj['confidence'] = confidences[i]result.append(result_obj)return {'result':result}if __name__ == '__main__':app.run(host='192.168.0.105')

客户端

import cv2
import requests
import base64def get_result(url,frame):retval, buffer = cv2.imencode('.jpg', frame)image = str(base64.b64encode(buffer), 'utf-8')json_data = {'img_b64': image}response = eval(requests.post(url, json=json_data).text).get('result')return response
if __name__ == '__main__':url = 'http://192.168.0.105:5000/predict'frame = cv2.imread('000003.jpg')response = get_result(url,frame)for result in response:confidence = result.get('confidence')x = int(result.get('x'))y = int(result.get('y'))w = int(result.get('w'))h = int(result.get('h'))name = result.get('name')frame = cv2.rectangle(frame, (x, y), (x + w, y + h), (255, 0, 0), 2)frame = cv2.putText(frame, name, (x, y), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)cv2.imshow('img',frame)cv2.waitKey(0)cv2.destroyAllWindows()

请求结果

2.3  摄像头识别

原理是图像识别+多线程

import cv2
import numpy as np
import queue
import threading
import timeLABELS = open("classes.names").read().strip().split("\n")
net = cv2.dnn.readNetFromDarknet('yolov4-tiny.cfg', 'yolov4-tiny_final.weights')
layer = net.getUnconnectedOutLayersNames()cap = cv2.VideoCapture(0)frame_queue = queue.Queue()
detection_result_queue = queue.Queue(maxsize=1)def video_capture():while cap.isOpened():ret, frame = cap.read()if ret:frame_queue.put(frame)cap.release()def predict():while cap.isOpened():start_time = time.time()predict_frame = frame_queue.get()(H, W) = predict_frame.shape[:2]blob = cv2.dnn.blobFromImage(predict_frame, 1 / 255.0, (416, 416), swapRB=True, crop=False)net.setInput(blob)layerOutputs = net.forward(layer)boxes = []confidences = []classIDs = []for output in layerOutputs:for detection in output:scores = detection[5:]classID = np.argmax(scores)confidence = scores[classID]box = detection[0:4] * np.array([W, H, W, H])(centerX, centerY, width, height) = box.astype("int")x = int(centerX - (width / 2))y = int(centerY - (height / 2))boxes.append([x, y, int(width), int(height)])confidences.append(float(confidence))classIDs.append(classID)idxs = cv2.dnn.NMSBoxes(boxes, confidences, 0.5, 0.3)if len(idxs) > 0:result_list = []for i in idxs.flatten():result_dic = {}(x, y) = (boxes[i][0], boxes[i][1])(w, h) = (boxes[i][2], boxes[i][3])label = LABELS[classIDs[i]]confidence = confidences[i]result_dic.__setitem__('x',x)result_dic.__setitem__('y',y)result_dic.__setitem__('w',w)result_dic.__setitem__('h',h)result_dic.__setitem__('label',label)result_dic.__setitem__('confidence',confidence)result_list.append(result_dic)detection_result_queue.put(result_list)print(time.time()-start_time)cap.release()def draw():while cap.isOpened():draw_frame = frame_queue.get()try:predict_results = detection_result_queue.get(block=False)for predict_result in predict_results:x = predict_result.get('x')y = predict_result.get('y')w = predict_result.get('w')h = predict_result.get('h')label = predict_result.get('label')confidence = predict_result.get('confidence')cv2.rectangle(draw_frame, (x, y), (x + w, y + h), (0,255,0), 1, lineType=cv2.LINE_AA)text = "{}: {:.4f}".format(label, confidence)cv2.putText(draw_frame, text, (x, y - 5), cv2.FONT_HERSHEY_SIMPLEX,0.5, (255,0,0), 1, lineType=cv2.LINE_AA)except:passcv2.imshow('draw_frame',draw_frame)cv2.waitKey(1)cap.release()threading.Thread(target=video_capture).start()
threading.Thread(target=predict).start()
threading.Thread(target=draw).start()

由于是CPU运行,效果根据CPU的性能会有差异,我是用NX上的6核CPU运行yolov4-tiny,大约0.5s一张。这个速度如果在摄像头上的连续识别就非常慢了,至少要1秒30帧才能有流畅的感觉,也就是0.03秒需要预测一张

  • 0.5/0.03 = 16.6 或许搞个17线程能够达成流畅的效果,没尝试过

3  darknet使用模型

优点:可以使用GPU

缺点:需要编译darknet才能使用

下面在代码中import darknet指的是import下面这个文件。下面这个py文件依赖了darknet文件夹下的其他文件,所以建议把下面的代码放在darknet的根目录下使用

3.1  识别图像

在模型读取的时候是通过data文件找names文件,而不是直接找names文件

import cv2
import darknet
import timenetwork, class_names, class_colors = darknet.load_network('/home/suyu/darknet/yolo_opencv/yolov7-tiny.cfg','/home/suyu/darknet/custom_training/custom_training.data','/home/suyu/darknet/yolo_opencv/yolov7-tiny_final_origin.weights',batch_size=1
)width = darknet.network_width(network)
height = darknet.network_height(network)
darknet_image = darknet.make_image(width, height, 3)image = cv2.imread('/home/suyu/darknet/yolo_opencv/280.jpg')
image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image_resized = cv2.resize(image_rgb, (width, height),interpolation=cv2.INTER_LINEAR)
darknet.copy_image_from_bytes(darknet_image, image_resized.tobytes())
start_time = time.time()
detections = darknet.detect_image(network, class_names, darknet_image, thresh=0.8)
print(time.time()-start_time)darknet.free_image(darknet_image)
image = darknet.draw_boxes(detections, image_resized, class_colors)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)cv2.imshow('image',image)
cv2.waitKey(0)
cv2.destroyAllWindows()

在终端上可以看到对模型进行了读取

画框的颜色每次都是随机的

在预测中,如果出现了下面这种情况,我们就需要减少预测的结果,可以通过标签后面的置信度减少,但是我们可以看到有两个框的置信度都在0.99以上,这个时候我们就需要用到其他减少框的办法

darknet.detect_image()这个方法,除了通过置信度减少预测结果,还可以通过非最大值抑制消除冗余,也就是最后一个的nms

  • hier_thresh也是消除冗余的一个参数,叫控制层次性阈值(与nms类似),没用过,如果nms不行的化再尝试使用它

nms数值越小,消除冗余的效果就越好,我这里直接改成了0.05

改完之后可以得到还不错的效果

3.2  摄像头识别

原理是识别图像+多线程,我用yolov7-tiny大概能到0.03秒预测一张,这个预测速度不加多线程应该也可以

import cv2
import numpy as np
import queue
import threading
import time
import darknetnetwork, class_names, class_colors = darknet.load_network('/home/suyu/darknet/yolo_opencv/yolov7-tiny.cfg','/home/suyu/darknet/custom_training/custom_training.data','/home/suyu/darknet/yolo_opencv/yolov7-tiny_final_origin.weights',batch_size=1
)width = darknet.network_width(network)
height = darknet.network_height(network)
darknet_image = darknet.make_image(width, height, 3)cap = cv2.VideoCapture(0)frame_queue = queue.Queue()
detection_result_queue = queue.Queue(maxsize=1)def video_capture():while cap.isOpened():ret, frame = cap.read()if ret:image_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)image_resized = cv2.resize(image_rgb, (width, height), interpolation=cv2.INTER_LINEAR)frame_queue.put(image_resized)cap.release()def predict():while cap.isOpened():start_time = time.time()image_resized = frame_queue.get()darknet.copy_image_from_bytes(darknet_image, image_resized.tobytes())detections = darknet.detect_image(network, class_names, darknet_image, thresh=0.8)# darknet.free_image(darknet_image)print(time.time()-start_time)detection_result_queue.put(detections)cap.release()def draw():while cap.isOpened():draw_frame = frame_queue.get()# print(draw_frame)try:detections = detection_result_queue.get(block=False)draw_frame = darknet.draw_boxes(detections, draw_frame, class_colors)except:passdraw_frame = cv2.cvtColor(draw_frame, cv2.COLOR_BGR2RGB)cv2.imshow('draw_frame',draw_frame)cv2.waitKey(1)cap.release()threading.Thread(target=video_capture).start()
threading.Thread(target=predict).start()
threading.Thread(target=draw).start()

如果你想对预测的结果进行别的操作你可以用到detection_result_queue中的detections,打印出来是这样的

列表套元组,元组的第一个值是label,第二个值是置信

操作的时候建议多开一个线程,然后将detection_result_queue置为2或者更高,避免两个线程抢数据的情况

4  onnx使用模型

onnx(Open Neural Network eXchange) 开放式神经网络交换,好多模型都可以转换为onnx类型的模型,相当于是人工智能模型界的docker了。以高适配性而著名。在性能上并没有优于其他模型。

4.1  用到的库

onnx,这个是将其他类型的模型转换为onnx类型的库,在arm端上安装可能会有些麻烦,但在amd上直接用pip就可以安装了

onnxruntime是跑onnx模型用的,onnxruntime在arm端可以直接用pip安装

4.2  darknet的weight转onnx

参考 https://zhuanlan.zhihu.com/p/543345367

源码 GitHub - Tianxiaomo/pytorch-YOLOv4: PyTorch ,ONNX and TensorRT implementation of YOLOv4

安装完onnx与onnxruntime后运行demo_darknet2onnx.py,第一个参数是cfg,第二个参数是names,第三个参数是weights,第四个参数是图,第五个参数是batch_size,直接写1就行了

python demo_darknet2onnx.py /home/suyu/darknet/cfg/yolov4.cfg /home/suyu/darknet/data/coco.names /home/suyu/darknet/yolov4.weights /home/suyu/darknet/data/dog.jpg 1

如果opencv版本过高会爆出下面两个问题,是cv2.rectangele()与cv2.putText()的参数需为int

我们需要更改 /pytorch-YOLOv4-master/tool/utils.py

把画红线的地方改成int

成功使用会显示下面这些东西

在文件夹中可以找到通过onnx预测成功的图像和onnx模型

4.3  识别图像

识别的代码是用的上面提供的源码中的东西,有的代码直接从utils中复制过来了,用的时候不需要引入其他py文件了,只需要.onnx文件与.names文件

import sys
import onnx
import os
import argparse
import numpy as np
import cv2
import onnxruntime
import time
import mathdef load_class_names(namesfile):class_names = []with open(namesfile, 'r') as fp:lines = fp.readlines()for line in lines:line = line.rstrip()class_names.append(line)return class_namesdef nms_cpu(boxes, confs, nms_thresh=0.5, min_mode=False):# print(boxes.shape)x1 = boxes[:, 0]y1 = boxes[:, 1]x2 = boxes[:, 2]y2 = boxes[:, 3]areas = (x2 - x1) * (y2 - y1)order = confs.argsort()[::-1]keep = []while order.size > 0:idx_self = order[0]idx_other = order[1:]keep.append(idx_self)xx1 = np.maximum(x1[idx_self], x1[idx_other])yy1 = np.maximum(y1[idx_self], y1[idx_other])xx2 = np.minimum(x2[idx_self], x2[idx_other])yy2 = np.minimum(y2[idx_self], y2[idx_other])w = np.maximum(0.0, xx2 - xx1)h = np.maximum(0.0, yy2 - yy1)inter = w * hif min_mode:over = inter / np.minimum(areas[order[0]], areas[order[1:]])else:over = inter / (areas[order[0]] + areas[order[1:]] - inter)inds = np.where(over <= nms_thresh)[0]order = order[inds + 1]return np.array(keep)def post_processing(img, conf_thresh, nms_thresh, output):# anchors = [12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401]# num_anchors = 9# anchor_masks = [[0, 1, 2], [3, 4, 5], [6, 7, 8]]# strides = [8, 16, 32]# anchor_step = len(anchors) // num_anchors# [batch, num, 1, 4]box_array = output[0]# [batch, num, num_classes]confs = output[1]t1 = time.time()if type(box_array).__name__ != 'ndarray':box_array = box_array.cpu().detach().numpy()confs = confs.cpu().detach().numpy()num_classes = confs.shape[2]# [batch, num, 4]box_array = box_array[:, :, 0]# [batch, num, num_classes] --> [batch, num]max_conf = np.max(confs, axis=2)max_id = np.argmax(confs, axis=2)t2 = time.time()bboxes_batch = []for i in range(box_array.shape[0]):argwhere = max_conf[i] > conf_threshl_box_array = box_array[i, argwhere, :]l_max_conf = max_conf[i, argwhere]l_max_id = max_id[i, argwhere]bboxes = []# nms for each classfor j in range(num_classes):cls_argwhere = l_max_id == jll_box_array = l_box_array[cls_argwhere, :]ll_max_conf = l_max_conf[cls_argwhere]ll_max_id = l_max_id[cls_argwhere]keep = nms_cpu(ll_box_array, ll_max_conf, nms_thresh)if (keep.size > 0):ll_box_array = ll_box_array[keep, :]ll_max_conf = ll_max_conf[keep]ll_max_id = ll_max_id[keep]for k in range(ll_box_array.shape[0]):bboxes.append([ll_box_array[k, 0], ll_box_array[k, 1], ll_box_array[k, 2], ll_box_array[k, 3], ll_max_conf[k], ll_max_conf[k], ll_max_id[k]])bboxes_batch.append(bboxes)t3 = time.time()print('-----------------------------------')print('       max and argmax : %f' % (t2 - t1))print('                  nms : %f' % (t3 - t2))print('Post processing total : %f' % (t3 - t1))print('-----------------------------------')return bboxes_batchdef plot_boxes_cv2(img, boxes, savename=None, class_names=None, color=None):import cv2img = np.copy(img)colors = np.array([[1, 0, 1], [0, 0, 1], [0, 1, 1], [0, 1, 0], [1, 1, 0], [1, 0, 0]], dtype=np.float32)def get_color(c, x, max_val):ratio = float(x) / max_val * 5i = int(math.floor(ratio))j = int(math.ceil(ratio))ratio = ratio - ir = (1 - ratio) * colors[i][c] + ratio * colors[j][c]return int(r * 255)width = img.shape[1]height = img.shape[0]for i in range(len(boxes)):box = boxes[i]x1 = int(box[0] * width)y1 = int(box[1] * height)x2 = int(box[2] * width)y2 = int(box[3] * height)bbox_thick = int(0.6 * (height + width) / 600)if color:rgb = colorelse:rgb = (255, 0, 0)if len(box) >= 7 and class_names:cls_conf = box[5]cls_id = box[6]print('%s: %f' % (class_names[cls_id], cls_conf))classes = len(class_names)offset = cls_id * 123457 % classesred = get_color(2, offset, classes)green = get_color(1, offset, classes)blue = get_color(0, offset, classes)if color is None:rgb = (red, green, blue)msg = str(class_names[cls_id])+" "+str(round(cls_conf,3))t_size = cv2.getTextSize(msg, 0, 0.7, thickness=bbox_thick // 2)[0]c1, c2 = (x1,y1), (x2, y2)c3 = (c1[0] + t_size[0], c1[1] - t_size[1] - 3)cv2.rectangle(img, (int(x1),int(y1)), (int(np.float32(c3[0])), int(np.float32(c3[1]))), rgb, -1)img = cv2.putText(img, msg, (int(c1[0]), int(np.float32(c1[1] - 2))), cv2.FONT_HERSHEY_SIMPLEX,0.7, (0,0,0), bbox_thick//2,lineType=cv2.LINE_AA)img = cv2.rectangle(img, (int(x1), int(y1)), (int(x2), int(y2)), rgb, bbox_thick)if savename:print("save plot results to %s" % savename)cv2.imwrite(savename, img)return imgdef detect(session, image_src, namesfile):IN_IMAGE_H = session.get_inputs()[0].shape[2]IN_IMAGE_W = session.get_inputs()[0].shape[3]# Inputresized = cv2.resize(image_src, (IN_IMAGE_W, IN_IMAGE_H), interpolation=cv2.INTER_LINEAR)img_in = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB)img_in = np.transpose(img_in, (2, 0, 1)).astype(np.float32)img_in = np.expand_dims(img_in, axis=0)img_in /= 255.0#print("Shape of the network input: ", img_in.shape)# Computeinput_name = session.get_inputs()[0].nameoutputs = session.run(None, {input_name: img_in})boxes = post_processing(img_in, 0.4, 0.6, outputs)class_names = load_class_names(namesfile)return plot_boxes_cv2(image_src, boxes[0], class_names=class_names)if __name__ == '__main__':session = onnxruntime.InferenceSession('yolov4_1_3_640_640_static.onnx')namesfile = 'classes.names'image_src = cv2.imread('3.png')detected_img = detect(session, image_src, namesfile)cv2.imshow('detected_img',detected_img)cv2.waitKey(0)cv2.destroyAllWindows()

4.4  摄像头识别

用的采集、预测、画 三线程。模型输入width与height都为320,两分类,在树莓派4B中大概是0.5s一张,大概是1秒2个识别帧

import cv2
import numpy as np
import queue
import threading
import timeimport sys
import onnx
import os
import argparse
import onnxruntime
import mathdef load_class_names(namesfile):class_names = []with open(namesfile, 'r') as fp:lines = fp.readlines()for line in lines:line = line.rstrip()class_names.append(line)return class_namesdef nms_cpu(boxes, confs, nms_thresh=0.5, min_mode=False):x1 = boxes[:, 0]y1 = boxes[:, 1]x2 = boxes[:, 2]y2 = boxes[:, 3]areas = (x2 - x1) * (y2 - y1)order = confs.argsort()[::-1]keep = []while order.size > 0:idx_self = order[0]idx_other = order[1:]keep.append(idx_self)xx1 = np.maximum(x1[idx_self], x1[idx_other])yy1 = np.maximum(y1[idx_self], y1[idx_other])xx2 = np.minimum(x2[idx_self], x2[idx_other])yy2 = np.minimum(y2[idx_self], y2[idx_other])w = np.maximum(0.0, xx2 - xx1)h = np.maximum(0.0, yy2 - yy1)inter = w * hif min_mode:over = inter / np.minimum(areas[order[0]], areas[order[1:]])else:over = inter / (areas[order[0]] + areas[order[1:]] - inter)inds = np.where(over <= nms_thresh)[0]order = order[inds + 1]return np.array(keep)def post_processing(img, conf_thresh, nms_thresh, output):box_array = output[0]confs = output[1]if type(box_array).__name__ != 'ndarray':box_array = box_array.cpu().detach().numpy()confs = confs.cpu().detach().numpy()num_classes = confs.shape[2]box_array = box_array[:, :, 0]max_conf = np.max(confs, axis=2)max_id = np.argmax(confs, axis=2)bboxes_batch = []for i in range(box_array.shape[0]):argwhere = max_conf[i] > conf_threshl_box_array = box_array[i, argwhere, :]l_max_conf = max_conf[i, argwhere]l_max_id = max_id[i, argwhere]bboxes = []for j in range(num_classes):cls_argwhere = l_max_id == jll_box_array = l_box_array[cls_argwhere, :]ll_max_conf = l_max_conf[cls_argwhere]ll_max_id = l_max_id[cls_argwhere]keep = nms_cpu(ll_box_array, ll_max_conf, nms_thresh)if (keep.size > 0):ll_box_array = ll_box_array[keep, :]ll_max_conf = ll_max_conf[keep]ll_max_id = ll_max_id[keep]for k in range(ll_box_array.shape[0]):bboxes.append([ll_box_array[k, 0], ll_box_array[k, 1], ll_box_array[k, 2], ll_box_array[k, 3], ll_max_conf[k], ll_max_conf[k], ll_max_id[k]])bboxes_batch.append(bboxes)return bboxes_batchdef get_color(c, x, max_val):colors = np.array([[1, 0, 1], [0, 0, 1], [0, 1, 1], [0, 1, 0], [1, 1, 0], [1, 0, 0]], dtype=np.float32)ratio = float(x) / max_val * 5i = int(math.floor(ratio))j = int(math.ceil(ratio))ratio = ratio - ir = (1 - ratio) * colors[i][c] + ratio * colors[j][c]return int(r * 255)def detect(session, image_src, namesfile):IN_IMAGE_H = session.get_inputs()[0].shape[2]IN_IMAGE_W = session.get_inputs()[0].shape[3]# Inputresized = cv2.resize(image_src, (IN_IMAGE_W, IN_IMAGE_H), interpolation=cv2.INTER_LINEAR)img_in = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB)img_in = np.transpose(img_in, (2, 0, 1)).astype(np.float32)img_in = np.expand_dims(img_in, axis=0)img_in /= 255.0# Computeinput_name = session.get_inputs()[0].nameoutputs = session.run(None, {input_name: img_in})boxes = post_processing(img_in, 0.4, 0.6, outputs)return boxes[0]session = onnxruntime.InferenceSession('yolov4_1_3_640_640_static.onnx')
namesfile = 'classes.names'
class_names = load_class_names(namesfile)cap = cv2.VideoCapture(0)frame_queue = queue.Queue()
detection_result_queue = queue.Queue(maxsize=1)def video_capture():while cap.isOpened():ret, frame = cap.read()if ret:frame_queue.put(frame)cap.release()def predict():while cap.isOpened():start_time = time.time()predict_frame = frame_queue.get()result_list = detect(session, predict_frame, namesfile)detection_result_queue.put(result_list)print(time.time()-start_time)cap.release()def draw():while cap.isOpened():draw_frame = frame_queue.get()try:boxes = detection_result_queue.get(block=False)img = draw_framewidth = img.shape[1]height = img.shape[0]for i in range(len(boxes)):box = boxes[i]x1 = int(box[0] * width)y1 = int(box[1] * height)x2 = int(box[2] * width)y2 = int(box[3] * height)bbox_thick = int(0.6 * (height + width) / 600)rgb = (255, 0, 0)if len(box) >= 7 and class_names:cls_conf = box[5]cls_id = box[6]print('%s: %f' % (class_names[cls_id], cls_conf))classes = len(class_names)offset = cls_id * 123457 % classesred = get_color(2, offset, classes)green = get_color(1, offset, classes)blue = get_color(0, offset, classes)rgb = (red, green, blue)msg = str(class_names[cls_id])+" "+str(round(cls_conf,3))t_size = cv2.getTextSize(msg, 0, 0.7, thickness=bbox_thick // 2)[0]c1, c2 = (x1,y1), (x2, y2)c3 = (c1[0] + t_size[0], c1[1] - t_size[1] - 3)cv2.rectangle(img, (int(x1),int(y1)), (int(np.float32(c3[0])), int(np.float32(c3[1]))), rgb, -1)img = cv2.putText(img, msg, (int(c1[0]), int(np.float32(c1[1] - 2))), cv2.FONT_HERSHEY_SIMPLEX,0.7, (0,0,0), bbox_thick//2,lineType=cv2.LINE_AA)img = cv2.rectangle(img, (int(x1), int(y1)), (int(x2), int(y2)), rgb, bbox_thick)draw_frame = img#except Exception as e:#print(e)except:passcv2.imshow('draw_frame',draw_frame)cv2.waitKey(1)cap.release()threading.Thread(target=video_capture).start()
threading.Thread(target=predict).start()
threading.Thread(target=draw).start()

在树莓派4B上是下面这个效果

 

本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如若转载,请注明出处:http://www.rhkb.cn/news/375917.html

如若内容造成侵权/违法违规/事实不符,请联系长河编程网进行投诉反馈email:809451989@qq.com,一经查实,立即删除!

相关文章

SQLite 命令行客户端 + HTA 实现简易UI

SQLite 命令行客户端 HTA 实现简易UI SQLite 客户端.hta目录结构V2V3参考资料 仅用于探索可行性&#xff0c;就只实现了 SELECT。 SQLite 客户端.hta <!DOCTYPE html> <html> <head><meta http-equiv"Content-Type" content"text/html;…

StarRocks分布式元数据源码解析

1. 支持元数据表 https://github.com/StarRocks/starrocks/pull/44276/files 核心类&#xff1a;LogicalIcebergMetadataTable&#xff0c;Iceberg元数据表&#xff0c;将元数据的各个字段做成表的列&#xff0c;后期可以通过sql操作从元数据获取字段&#xff0c;这个表的组成…

Java版Flink使用指南——从RabbitMQ中队列中接入消息流

大纲 创建RabbitMQ队列新建工程新增依赖编码设置数据源配置读取、处理数据完整代码 打包、上传和运行任务测试 工程代码 在《Java版Flink使用指南——安装Flink和使用IntelliJ制作任务包》一文中&#xff0c;我们完成了第一个小型Demo的编写。例子中的数据是代码预先指定的。而…

人工智能大模型如何助力电商产品经理打造高效的商品工业属性画像

摘要 商品工业属性画像是电商产品经理在进行商品管理、推荐、搜索、广告等业务时的重要依据。通过对商品的工业属性&#xff08;如品类、品牌、规格、功能、风格等&#xff09;的准确识别和标注&#xff0c;可以提高商品的展示效果、匹配度、转化率和用户满意度。然而&#xf…

如何为IP申请SSL证书

目录 以下是如何轻松为IP地址申请SSL证书的详细步骤&#xff1a; 申请IP证书的基本条件&#xff1a; 申请IP SSL证书的方式&#xff1a; 确保网络通信安全的核心要素之一&#xff0c;是有效利用SSL证书来加密数据传输&#xff0c;特别是对于那些直接通过IP地址访问的资源。I…

Java 将图片转base64和base64转图片

工具 Base64 和 图片互转。 导入的依赖 <!-- https://mvnrepository.com/artifact/com.sun.xml.bind/jaxb-impl --><dependency><groupId>com.sun.xml.bind</groupId><artifactId>jaxb-impl</artifactId><version>4.0.5</versi…

【linux】进程间通信(IPC)——匿名管道,命名管道与System V内核方案的共享内存,以及消息队列和信号量的原理概述

目录 ✈必备知识 进程间通信概述 &#x1f525;概述 &#x1f525;必要性 &#x1f525;原理 管道概述 &#x1f525;管道的本质 &#x1f525;管道的相关特性 &#x1f525;管道的同步与互斥机制 匿名管道 &#x1f525;系统调用接口介绍 &#x1f525;内核原理 …

【开发环境】搭建PX4+ROS2+MAVROS2+Simulink+Optitrack实物联合仿真环境

搭建PX4ROS2MAVROS2SimulinkOptiTrack实物联合仿真环境 Ubuntu中的安装过程下载并编译PX4固件代码安装ROS2安装VRPN动捕数据转换ROS2话题库安装VRPN库拉取vrpn_client_ros2节点代码并配置VRPN server参数编译软件包启动vrpn_client_ros2节点重命名话题名 /vrpn/fly/pose 为 /ma…

Mac VSCode 突然闪退、崩溃、打不开了

vscode 1.90.2版本下载&#xff0c;刚上传还在审核中 1、 思路历程 VSCode 作为前端常用开发工具&#xff0c;其重要性就不一一描述了。 所以 VSCode 突然打不开了&#xff0c;真的是让我一脸懵逼。 本来以为问题不大&#xff0c;于是 &#xff1a; 1、重启了一下VSCode 2、…

Python 爬虫:使用打码平台来识别各种验证码:

本课程使用的是 超级鹰 打码平台&#xff0c; 没有账户的请自行注册&#xff01; 超级鹰验证码识别-专业的验证码云端识别服务,让验证码识别更快速、更准确、更强大 使用打码平台来攻破验证码难题&#xff0c; 是很简单容易的&#xff0c; 但是要钱&#xff01; 案例代码及测…

记录些Redis题集(1)

Redis内存淘汰触发条件的相关配置如下&#xff1a; Redis通过配置项maxmemory来设定其允许使用的最大内存容量。当Redis实际占用的内存达到这一阈值时&#xff0c;将触发内存淘汰机制&#xff0c;开始删除部分数据以释放内存空间&#xff0c;防止服务因内存溢出而异常。 Redi…

vitest 单元测试应用与配置

vitest 应用与配置 一、简介 Vitest 旨在将自己定位为 Vite 项目的首选测试框架&#xff0c;即使对于不使用 Vite 的项目也是一个可靠的替代方案。它本身也兼容一些Jest的API用法。 二、安装vitest // npm npm install -D vitest // yarn yarn add -D vitest // pnpm pnpm …

【Linux】vim详解

1.什么是vi/vim? 简单来说&#xff0c;vi是老式的文本编辑器&#xff0c;不过功能已经很齐全了&#xff0c;但是还是有可以进步的地方。vim则可以说是程序开发者的一项很好用的工具&#xff0c;就连 vim的官方网站&#xff08; http://www.vim.org&#xff09;自己也说vim是一…

【公益案例展】华为云X《无尽攀登》——攀登不停,向上而行

‍ 华为云公益案例 本项目案例由华为云投递并参与数据猿与上海大数据联盟联合推出的 #榜样的力量# 《2024中国数据智能产业最具社会责任感企业》榜单/奖项”评选。 大数据产业创新服务媒体 ——聚焦数据 改变商业 夏伯渝&#xff0c;中国无腿登珠峰第一人&#xff0c;一生43年…

语言模型演进:从NLP到LLM的跨越之旅

在人工智能的浩瀚宇宙中&#xff0c;自然语言处理&#xff08;NLP&#xff09;一直是一个充满挑战和机遇的领域。随着技术的发展&#xff0c;我们见证了从传统规则到统计机器学习&#xff0c;再到深度学习和预训练模型的演进。如今&#xff0c;我们站在了大型语言模型&#xff…

免费开源数字人生成工具

使用步骤更是简单到不行&#xff1a; 1. 输入图片&#xff1a;选择你想要生成动态视频的肖像图片。 2. 输入音频&#xff1a;提供与图片匹配的音频文件&#xff0c;EchoMimic会根据音频内容驱动肖像的动态效果。 3. 设置参数&#xff1a;一般保持默认设置即可&#xff0c;当然&…

端到端自动驾驶系列(一):自动驾驶综述解析

端到端自动驾驶系列(一)&#xff1a;自动驾驶综述解析 End-to-end-Autonomous-Driving Abstract Abstract—The autonomous driving community has witnessed a rapid growth in approaches that embrace an end-to-end algorithm framework, utilizing raw sensor input to …

模块化(一)nodejs

模块化 一.模块化的基本概念1.1 什么是模块化1.2 模块化规范 二.Node.js 中的模块化2.1 Node.js 中模块的分类2.2 加载模块2.3 Node.js 中的模块作用域2.4 向外共享模块作用域中的成员 一.模块化的基本概念 1.1 什么是模块化 模块化 是指解决一个 复杂问题 时&#xff0c;自顶…

uni-app 保存号码到通讯录

1、 添加模块 2、添加权限 3、添加策略 Android&#xff1a; "permissionExternalStorage" : {"request" : "none","prompt" : "应用保存运行状态等信息&#xff0c;需要获取读写手机存储&#xff08;系统提示为访问设备上的照片…

安卓14中Zygote初始化流程及源码分析

文章目录 日志抓取结合日志与源码分析systemServer zygote创建时序图一般应用 zygote 创建时序图向 zygote socket 发送数据时序图 本文首发地址 https://h89.cn/archives/298.html 最新更新地址 https://gitee.com/chenjim/chenjimblog 本文主要结合日志和代码看安卓 14 中 Zy…